As generative AI tools reshape the music industry in transformative—and occasionally controversial—ways, Google is significantly increasing its investment in AI technologies to craft original songs and lyrics. Today, the tech giant introduced MusicFX, an enhanced iteration of its previously launched MusicLM, a music-generating tool from last year. MusicFX can produce melodies lasting up to 70 seconds along with music loops, providing what Google describes as “higher-quality” and “faster” music generation.
MusicFX is now accessible through Google’s AI Test Kitchen app, which allows users to experiment with cutting-edge AI-powered innovations from the company's labs. While MusicFX initially launched for a select group of users in December, it is now broadly available.
So, how does it sound? After testing it briefly, I found MusicFX produced decent results. In fact, music generation tools are advancing to the point where distinguishing between outputs has become challenging. The current top-end models yield impressively crisp and clean tracks; however, they often lean towards being uninspired and melodically unfocused.
Feeling nostalgic, I used the prompt “a house music song with funky beats that’s danceable and uplifting, with summer rooftop vibes.” MusicFX delivered reasonably well, though the tracks didn’t quite measure up to some of the outstanding DJ sets I’ve recently experienced.
Listen for yourself:
-
-
On the other hand, stringed instruments sounded noticeably less impressive, resembling cheap MIDI samples—possibly due to MusicFX’s limited training data. Here’s a glimpse of two tracks generated from the prompt “a soulful melody played on string instruments, orchestral, with a strong melodic core”:
-
-
For variety, here’s MusicFX’s rendition of “a weepy song on guitar, melancholic, slow tempo, on a moonlight night” (note the misspelling):
-
-
It's important to emphasize that MusicFX has limitations—specifically, it avoids generating tracks that mention specific artists or contain vocals, to navigate copyright issues. Google employs SynthID, a watermarking technology created by its DeepMind division, to identify tracks produced by MusicFX.
While I initially struggled to grasp the exact filtering list Google uses for artist and song name censorship, I managed to bypass some restrictions. Although prompts referencing iconic artists like SZA and The Beatles were rejected, my requests involving Lake Street Dive were accepted, even if the final outputs weren't remarkable.
On the lyrical side, Google has unveiled TextFX, a new lyrics generator intended to accompany MusicFX. Like MusicFX, TextFX has been quietly available to a limited audience but is now more widely accessible and has received upgrades for improved user experience and navigation.
According to Google, TextFX was developed in collaboration with rapper and producer Lupe Fiasco and operates on PaLM 2, one of the company’s text-generating AI models. The tool "draws inspiration from the lyrical and linguistic techniques Fiasco has cultivated throughout his career."
I initially thought TextFX would operate solely as an automated lyrics generator; however, it proves to be far more complex. It comprises various modules designed to assist in the lyric-writing process, including features that identify words starting with a specified letter and find correlations between unrelated ideas.
Though TextFX takes some time to master, I believe it will prove beneficial for songwriters and creatives alike. However, users should examine its outputs carefully. Google warns that TextFX “may display inaccurate information, including about individuals.” I found that it could generate misleading claims, such as suggesting that climate change “is a hoax perpetrated by the Chinese government to undermine American businesses.”
Remaining Questions
With MusicFX and TextFX, Google demonstrates its substantial commitment to generative AI music technology. However, I question whether its focus on competition will enable it to effectively confront the complex ethical issues surrounding AI-generated music in the long run.
Increasingly, DIY tracks using generative AI to recreate familiar sounds and vocal patterns are gaining viral traction. Music labels have quickly flagged AI-generated music to streaming platforms like Spotify and SoundCloud due to intellectual property concerns, generally winning those disputes. Yet, uncertainties persist regarding the legal implications of “deepfake” music and its potential copyright violations involving artists, labels, and rights holders.
In August, a federal ruling established that AI-generated art does not qualify for copyright protection. However, the U.S. Copyright Office has yet to take a definitive stance and has only recently started soliciting public feedback on AI-related copyright matters. It's still unclear if users could face legal repercussions for attempting to monetize music created in the style of established artists.
Google is navigating this sensitive terrain thoughtfully by testing AI models, developed in partnership with artists such as Alec Benjamin, Charlie Puth, Charli XCX, Demi Lovato, John Legend, Sia, and T-Pain, for applications on YouTube. This thoughtful approach contrasts with some of Google’s competitors, like Stability AI, which contends that “fair use” justifies using content for training without creator permission.
However, with music labels pursuing cases against generative AI companies over copyrighted lyrics in training datasets and artists voicing their dissatisfaction, Google faces significant challenges—and it shows no signs of slowing down.
Ranking
